What Is the Jev Model? TypeSafe System One API Guide
The Jev model that developers have been discussing recently is not a general chat model for writing, code generation, or long-form answers. It is TypeSafe AI's first System One Model: your application supplies state and focused questions, Jev returns typed decisions such as Choice, Score, or Noul with probabilities and confidence, and your code decides whether to route, block, escalate, or call another model.
TypeSafe announced Jev as an early-access product on September 15, 2026. This guide turns the official model reference, API shape, and practical boundaries into an integration checklist. Verify access, price, and aliases in the official model documentation; this site's gateway availability and billing are separate and belong to the live model pricing page.
How is Jev different from a regular LLM?
Most LLMs primarily return strings: answers, code, summaries, or explanations. Jev returns a shape defined by the caller, so the model answers the judgment that the software actually needs.
| Comparison | General LLM | Jev / System One |
|---|---|---|
| Input | Conversation, prompt, and context | state plus typed questions |
| Output | Free text that must be parsed and validated | Structured Choice, Score, or Noul results |
| Best fit | Writing, chat, code, research | Classification, routing, scoring, gates |
| Uncertainty | Often requires extra prompting | Probabilities or confidence can accompany decisions |
| Next action | A person or agent interprets text | Application code applies thresholds and rules |
Think of Jev as a probabilistic judgment function: State → Question → Decision → Action. Jev contributes the decision signal, not the business action. Refunds, account blocks, deletions, and external messages still require explicit rules and an approval path.
Which Jev model IDs are available?
TypeSafe's model documentation lists one pinned version and two aliases:
| Model ID | Meaning | Best use |
|---|---|---|
jev-latest |
Alias for the current stable release | Development, evaluation, and following stable updates |
jev-preview |
Preview alias; currently shown as the same target as jev-latest |
Only when preview behavior is explicitly needed |
jev-1.13.0 |
Pinned version ID | Production, threshold regression, and reproducible tests |
The official page currently shows both jev-latest and jev-preview pointing to jev-1.13.0, and warns that aliases can move when a new release ships. If your workflow depends on tuned thresholds, log the response's actual model field or pin a version so an update cannot silently change routing results.
How do I make a Jev API request?
The official endpoint is POST https://api.typesafe.ai/v1/systemone. A request contains state, model, and questions. One call can evaluate Choice, Score, and Noul questions in parallel:
curl -X POST https://api.typesafe.ai/v1/systemone \
-H "Authorization: Bearer $TYPESAFE_API_KEY" \
-H "Content-Type: application/json" \
-d @- <<'EOF'
{
"state": "A customer says a payment has failed for three days and the order is about to expire.",
"model": "jev-latest",
"questions": {
"department": {
"type": "choice",
"instructions": "Which team should handle this?",
"criteria": {
"billing": "Payment, order, or subscription issue",
"technical": "API or integration failure",
"support": "General question or human follow-up"
}
},
"urgency": {
"type": "score",
"instructions": "How urgent is this message?",
"criteria": ["Routine", "Urgent", "Critical"]
},
"needs_human": {
"type": "noul",
"instructions": "Should this request be escalated to a human?"
}
}
}
EOF
The response contains typed answers rather than natural language that needs regex parsing. Choice can include a selected option and probability distribution; Score can include a score and probabilities; Noul returns a yes/no judgment and uncertainty. Record the model version, question definitions, confidence, and final action so the decision loop can be replayed and its thresholds improved.
When should I use Choice, Score, or Noul?
- Choice: select one category from a finite set, such as a support queue, risk type, or next tool.
- Score: rate an ordered dimension, such as urgency, relevance, or sentiment.
- Noul: assign a probability to a clear yes/no question, such as whether human review is needed.
Keep questions atomic. Do not ask one question to classify a ticket, score sentiment, assess risk, and write a response. Split those judgments into typed questions and combine the results in code so they can be tested independently.
What production workflows fit Jev?
Jev fits decisions that change a software branch:
- Support-ticket routing and urgency scoring;
- Content moderation, risk gates, and human-review triggers;
- Agent tool selection, request routing, and next-step branches;
- Document relevance, candidate ranking, and retrieval filtering;
- Structured checks on another model's output before continuing a workflow.
A practical architecture is: a general LLM understands, writes, or proposes; Jev makes the bounded decision; application code owns permissions, thresholds, retries, and the final action. Do not let a low-confidence result directly issue a refund, delete data, or send an external notification.
What is Jev not good at?
Jev does not generate free-form text. It should not replace a writing model, translation model, coding model, or long-form research agent. It also does not replace exact arithmetic, permission systems, or audit rules.
The current official model page says Jev accepts text or text-shaped state and does not directly accept images, audio, or video. Pre-process non-text material into text or structured fields with another tool. English is the primary training language; validate confidence and error patterns on your own non-English samples before using Jev for a critical workflow.
How should I evaluate speed and price?
The official model page currently lists a 64K context, input-token billing, and free output tokens. TypeSafe's published Jev 1.13 price is $0.042 / 1M input tokens. These are provider facts, not this gateway's final price, and they do not guarantee the same rate limits for every account.
TypeSafe also says rate limits can change while demand is being served. A useful evaluation records the question definitions, input length, concurrency, end-to-end latency, confidence, human-review rate, and final business accuracy—not just one “fast” or “cheap” screenshot.
What should I verify before using Jev through this site?
This site's model catalog exposes Jev routes and live billing status. Use the exact Model ID shown by the live catalog, send a low-risk request first, and verify the protocol, response shape, and billing before expanding traffic. A third-party gateway's endpoint, aliases, and billing are not the same contract as the vendor's direct API.
Use GPT, Claude, DeepSeek, Qwen, or another generative model when the workflow needs text, code, or open-ended agent work. Put Jev at the decision node when the workflow needs routing, scoring, or a gate.
FAQ
Is Jev a large language model?
Not in the usual text-generation sense. Jev is TypeSafe AI's System One model: it receives state and typed questions and returns structured decisions, probabilities, and confidence.
Can Jev write, translate, or generate code?
No. Its output types are defined by the questions. Use a generative model for writing, translation, or code generation.
Should I use jev-latest or jev-1.13.0?
Use jev-latest while developing if you want the current stable alias. Pin jev-1.13.0 in production when thresholds and reproducibility matter, then run regression checks before upgrading.
Does Jev's probability mean it is always correct?
No. Probability and confidence are decision signals, not a business guarantee. Keep a human-review band and measure calibration, error types, and drift on labeled samples.
Can Jev replace my agent?
No. Jev is a structured decision node inside an agent workflow. Generative models, tools, permissions, and application code still have their own jobs.
Sources
- TypeSafe AI model documentation: model IDs, aliases, context, input limits, and pricing.
- TypeSafe AI Quick Start: request shape, Choice/Score/Noul examples, and SDKs.
- TypeSafe AI launch article: System One and Jev positioning, launch details, and vendor claims.
- TypeSafe AI Jev overview: an independent developer-facing explanation of structured decisions and boundaries.
- This site's live model pricing: gateway routes, billing, and availability.
Sources checked on October 1, 2026. This is an integration and selection guide, not an independent benchmark.